JP6856503B2 - Impression estimation model learning device, impression estimation device, impression estimation model learning method, impression estimation method, and program - Google Patents

Description

The present invention relates to a technique for automatically estimating the impression of data.

There is a need for technology that automatically estimates the impression of data such as voice and images. For example, if the impression (brightness, lightness, etc.) of music or an image can be estimated, the impression value of the data is automatically added and can be used for data search. In addition, if the impression (naturalness, likability, etc.) of the call center operator can be estimated, the response quality of the operator can be automatically evaluated.

Automatic estimation of the impression of data using machine learning can be realized by manually creating data to which impression values are given in advance and learning a machine learning model that links the data and impression values. At this time, since the evaluation criteria of the impression value given manually differ depending on the person, it is necessary to prepare the learning data as follows, for example, when learning the machine learning model (Non-Patent Document 1). reference).
1. 1. Impression values (several grades) are manually labeled by multiple people.
2. If necessary, the impression value of several steps is rounded to a label of 2 to 3 steps.
3. 3. In order to exclude data that is blurred by people from the correct answer data, only labels that match with multiple people are selected (if three or more people give the label, the label may be selected by majority vote).

In order to learn a model that estimates a stable impression value, an evaluator who can give a stable impression value (for example, an evaluator who can give the same impression value when listening to the same data on different days) is selected. However, in order to give the impression value while stabilizing it, it is necessary to perform the impression value giving work with a break (see Non-Patent Document 2).

Masaki Sake Brewery et al., “Construction of Natural Speech Database for Emotion / Emotion Discrimination”, IPSJ Journal, vol. 52, No. 3, pp. 1185-1194, 2011 Takayuki Kagomiya et al., “Summary of Impression Rating Data”, [online], [Searched on November 1, 2017], Internet <URL: http://pj.ninjal.ac.jp/corpus_center/csj/manu- f / impression.pdf>

In order to learn a stable impression estimation model, it is necessary to select a stable evaluator, evaluate for a long period of time with a break, and exclude data with blurring, which is extremely costly. .. In addition, if an evaluator is selected to some extent, an impression estimation model specialized for the tendency of several evaluators is learned, and the estimated impression value is the impression value of the average feeling of the user. It may be different.

In view of the above points, an object of the present invention is to stably estimate an average impression value without selecting an evaluator or excluding data, taking into account the tendency of the evaluator's impression evaluation. It is to realize the impression estimation technology that can be done.

In order to solve the above problems, the impression estimation model learning device according to the first aspect of the present invention is an evaluation in which the learning data, the correct answer value of the impression value given to the learning data, and the correct answer value of the impression value are given. Evaluator tendency data storage that stores the evaluator tendency database that associates the learning data storage unit that stores the learning database associated with the person and the evaluator tendency that represents the appearance probability of each impression value in the evaluator and the evaluator. The unit, the impression estimation model storage unit that stores the impression estimation model that estimates the posterior probability of the impression value from the training data, the weight estimation model storage unit that stores the weight estimation model that estimates the weight from the evaluator tendency, and the evaluator. The impression estimation unit that estimates the posterior probability of the impression value from the training data using the impression estimation model for each group that classifies the tendency, and the weight that estimates the weight from the evaluator tendency associated with the training data using the weight estimation model. The weight estimation model update unit that updates the weight estimation model from the difference between the estimation unit and the posterior probability of the impression value weighted by the weight and the correct answer value of the impression value given to the learning data, and the weight estimation model update unit that updates the weight from the posterior probability of the impression value. It includes an impression estimation model update unit that updates the impression estimation model as a learning weight.

In order to solve the above problems, the impression estimation device of the second aspect of the present invention includes an impression estimation model storage unit that stores an impression estimation model learned by the impression estimation model learning device of the first aspect, and a first impression estimation model storage unit. Impression estimation model of the above aspect Using the weight estimation model storage unit that stores the weight estimation model learned by the learning device, and the impression estimation model for each group that classifies the evaluator tendency that represents the appearance probability for each impression value in the evaluator. An impression estimation unit that estimates the posterior probability of the impression value from the input data, a weight estimation unit that estimates the weight from the evaluator tendency input using the weight estimation model, and an integration that is the posterior probability of the impression value weighted by the weight. It includes an integrated value estimation unit that obtains the impression value of input data based on posterior probabilities.

According to the impression estimation technique of the present invention, it is not necessary for all evaluators to give impression values to all the training data, so that it is not necessary to select an evaluator or restraint for a long period of time, and it is not necessary to select a label. Therefore, the cost of creating learning data can be reduced. This makes it possible to stably estimate the average impression value for the impression value whose standard differs depending on the person. That is, according to the impression estimation technique of the present invention, it is possible to stably estimate the average impression value while taking into account the tendency of the evaluator's impression evaluation without selecting the evaluator or excluding the data. it can.

FIG. 1 is a diagram illustrating a functional configuration of an impression estimation model learning device. FIG. 2 is a diagram illustrating the processing procedure of the impression estimation model learning method. FIG. 3 is a diagram illustrating the functional configuration of the impression estimation device. FIG. 4 is a diagram illustrating the processing procedure of the impression estimation method.

The points of the present invention are as follows. First, the fluctuation of the impression value by the evaluator is automatically classified into a group for each tendency by a plurality of impression estimation models. Next, the results of automatic classification of the impression value estimation results are weighted according to the tendency of the impression value to fluctuate by the evaluator, and the results are integrated to estimate the final impression value. In this way, the average impression value is estimated by giving the average deviation of the evaluation by the evaluator to the estimation result of the impression value.

Hereinafter, embodiments of the present invention will be described in detail. In the drawings, the components having the same function are given the same number, and duplicate description will be omitted.

An embodiment of the present invention includes an impression estimation model learning device that learns an impression estimation model and a weight estimation model using training data and evaluator tendency data prepared in advance, and a trained impression estimation model and a weight estimation model. It is roughly divided into an impression estimation device that estimates the impression value of the input data using. Hereinafter, the details of the functional configuration and the processing procedure of the impression estimation model learning device and the impression estimation device of the embodiment will be described in order.

The symbol "^" etc. used in the text should be written directly above the character immediately after it, but due to restrictions on the text notation, it should be written immediately before the character. In the mathematical formula, these symbols are described in their original positions, that is, directly above the letters.

<Impression estimation model learning device>
As shown in FIG. 1, the impression estimation model learning device 1 of the embodiment includes N (≧ 2) impression estimation units 11-n (n = 1, 2, ..., N), weight estimation units 12, and weight estimation units 12. It includes a model learning unit 13, a learning data storage unit 14, an evaluator tendency data storage unit 15, and a weight estimation model storage unit 16. Each impression estimation unit 11-n includes an impression estimation model storage unit 111-n. The model learning unit 13 includes a weight estimation model update unit 131 and an impression estimation model update unit 132. The impression estimation model learning method of the embodiment is realized by the impression estimation model learning device 1 performing the processing of each step illustrated in FIG.

The impression estimation model learning device 1 is configured by loading a special program into a known or dedicated computer having, for example, a central processing unit (CPU), a main storage device (RAM: Random Access Memory), or the like. It is a special device. The impression estimation model learning device 1 executes each process under the control of the central processing unit, for example. The data input to the impression estimation model learning device 1 and the data obtained by each process are stored in, for example, the main storage device, and the data stored in the main storage device is read out to the central processing unit as needed. It is used for other processing. At least a part of each processing unit of the impression estimation model learning device 1 may be configured by hardware such as an integrated circuit. Each storage unit included in the impression estimation model learning device 1 is, for example, an auxiliary storage device composed of a main storage device such as RAM (Random Access Memory) and a semiconductor memory element such as a hard disk, an optical disk, or a flash memory (Flash Memory). , Or it can be configured with middleware such as a relational database or key-value store.

The learning data storage unit 14 stores a learning database that stores learning data for learning the impression estimation model. An example of the learning database is shown in Table 1. As shown in Table 1, the learning database has the training data x (j), the impression value y (j), and the evaluator number i (for the data number j (j = 0, 1,…, J-1). j) is associated. Here, J is the total number of training data.

The learning data x (j) indicates data such as images and sounds. Since a plurality of evaluators give impression values to one learning data x (j), there are a plurality of learning data x (j) having the same value in the learning database (for example, j = 0 in Table 1). And j = 1, j = 2, j = 3, and j = 4, etc.). The training data x (j) is, for example, a known feature amount extracted from image data or audio data.

The impression value y (j) is an impression value manually given, that is, a correct answer value of the impression value. The impression value y (j) is a vector having the dimension of the number of steps to be evaluated, and is a one-hot vector in which the dimension of the evaluated value is 1 and the other dimensions are 0. For example, when "brightness" is evaluated on a scale of "very dark: 1", "dark: 2", "normal: 3", "bright: 4", and "very bright: 5", "very dark: 1". When evaluated as [1, 0, 0, 0, 0], when evaluated as "dark: 2", [0, 1, 0, 0, 0], when evaluated as "normal: 3" [0, 0, 1, 0, 0], [0, 0, 0, 1, 0] when evaluated as "bright: 4", [0, 0] when evaluated as "very bright: 5" , 0, 0, 1].

The evaluator number i (j) is an identifier (ID) indicating an evaluator who has given the impression value y (j) to the learning data x (j). The evaluator number i (j) is a number associated with the evaluator number i of the evaluator tendency database stored in the evaluator tendency data storage unit 15.

In normal impression estimation, all evaluators give impression values to a certain data, estimate the average value as a label of training data, or eliminate inconsistent impression values for learning. However, in reality, it requires a huge cost for all evaluators to label all the data and obtain a sufficient amount of training data. Therefore, in the present embodiment, one or more evaluators give an impression value to a certain data, and the tendency of the evaluator is given as auxiliary information for learning. This makes it possible to learn the impression value of a certain data without requiring evaluation by everyone.

The evaluator tendency data storage unit 15 stores an evaluator tendency database accumulating the evaluator tendency, which is auxiliary information given when learning the impression estimation model. Table 2 shows an example of the evaluator tendency database. In the evaluator tendency database, the evaluator tendency q (i, k) (k = 0, 1,…, K-1) is associated with the evaluator number i (i = 0, 1,…, I-1). Has been done. Here, I is the total number of evaluators, and K is the number of types of evaluation values.

The evaluator tendency q (i, k) is a vector representing how much the evaluator i gives the impression value to which stage (in other words, the appearance probability for each impression value k in the evaluator i). For example, when "brightness" is evaluated on a scale of "very dark: 1", "dark: 2", "normal: 3", "bright: 4", and "very bright: 5", the evaluator with i = 0 1% rated "very dark: 1", 2% rated "dark: 2", 30% rated "normal: 3", 58% rated "bright: 4", and "very dark" The evaluation of "bright: 5" is evaluated with a distribution of 9%, and the evaluator tendency q (i, k) at that time is a vector such as [1%, 2%, 30%, 58%, 9%]. ..

Hereinafter, the impression estimation model learning method executed by the impression estimation model learning device 1 of the embodiment will be described with reference to FIG.

In step S11, the impression estimation unit 11-n uses the impression estimation model n stored in the impression estimation model storage unit 111-n to obtain posterior probabilities p (j, n) of the impression value with respect to the training data x (j). Is calculated and output. The impression estimation model n is a model that estimates the posterior probability of the impression value from the input data, and learns the evaluation tendency. For example, some evaluators evaluate an impression on {4, 5}, some evaluate on {3, 4, 5}, and some evaluate on {1, 2, 3, 4, 5}. There is also a group of evaluators who do. At this time, the impression estimation unit 11-1 estimates the impression with {4,5}, the impression estimation unit 11-2 estimates the impression with {3, 4, 5}, and the impression estimation unit 11-3 estimates the impression. Each impression estimation unit 11-n makes an impression estimation that reflects the tendency of the evaluator's group, such as estimating with {1, 2, 3, 4, 5}. The number N of the impression estimation unit 11-n may be clustered to automatically determine the number of clusters using the evaluator tendency database, and the obtained number of clusters may be N. As the clustering for automatically determining the number of clusters, X-means or the like described in Reference 1 can be used.

[Reference 1] Dan Pelleg, Andrew W. Moore, "X-means: Extending K-means with Efficient Estimation of the Number of Clusters", ICML 2000, pp. 727-734, 2000.

The evaluator tendency is

Therefore, q (i, k) is under the same conditions as the Dirichlet distribution (the sum of each dimension is always 1). Reference 1 calculates the Bayesian Information Criterion (BIC), which determines clusters by the k-means algorithm and determines whether to increase the number of clusters, assuming that the clusters follow a normal distribution. However, the number of clusters is calculated by the EM algorithm instead of k-means, and the number of groups of evaluator tendency is calculated more strictly by calculating the BIC based on the mixed diricre distribution. be able to.

The posterior probability indicates the probability that the training data x (j) is estimated at each stage. For example, when the posterior probability when evaluating "brightness" in the above five stages is p (j, n) = [0.10, 0.20, 0.30, 0.25, 0.15], the training data x (j) is "extremely". The probability of "dark: 1" is 0.1, the probability of "dark: 2" is 0.2, the probability of "normal: 3" is 0.30, the probability of "bright: 4" is 0.25, and the probability of "very bright: 5" is It becomes 0.15.

The impression estimation unit 11-n can be estimated by a general machine learning method such as a neural network or a support vector machine (SVM). In the impression estimation model n used in the impression estimation unit 11-n, parameters are randomly set at the initial learning value.

In step S12, the weight estimation unit 12 uses the weight estimation model stored in the weight estimation model storage unit 16 to weight from the evaluator tendency q (i, k) stored in the evaluator tendency data storage unit 15. Find α (j, n). The weight estimation model is a model that estimates weights from the input evaluator tendency. The weight α (j, n) is a value of a ratio indicating how much of the impression estimation units 11-1 to 11-N considers a certain impression estimation unit 11-n as the final estimation result. The weight estimation unit 12 can also obtain the weight α (j, n) as a posterior probability in the same manner as the impression estimation unit 11-n. For example, when the weights in the impression estimation units 11-1 to 11-3 are α (j, 1) = 0.1, α (j, 2) = 0.6, and α (j, 3) = 0.3, the impression estimation unit 11-1 Is 10%, impression estimation unit 11-2 is 60%, and impression estimation unit 11-3 is 30%, which will be considered as the final estimation result.

The weight estimation unit 12 can estimate by a general machine learning method such as a neural network or a support vector machine. In the weight estimation model used in the weight estimation unit 12, parameters are randomly set at the initial learning value.

In step S13, the model learning unit 13 stores the weight α (j, n) output by the weight estimation unit 12, the posterior probability p (j, n) output by the impression estimation unit 11-n, and the learning data storage unit 14. Using the learning data x (j) and the impression value y (j), the weight estimation model stored in the weight estimation model storage unit 16 and the impression stored in the impression estimation model storage unit 111-n. Update the estimation model n. The model learning unit 13 includes both the weight estimation model update unit 131 and the impression estimation model update unit 132, and the weight estimation model and the impression estimation model n are alternately updated when learning is repeated.

In step S131, the weight estimation model update unit 131 updates the weight estimation model when the posterior probabilities p (j, n) are given. The weight estimation model is modeled by, for example, a neural network or a support vector machine. Weight sum of impression value y (j) of training data x (j) and posterior probability p (j, n)

Update the weight estimation model from the difference with. When the weight estimation model is a neural network, the weight estimation model can be updated from the general learning method (error propagation method) of the neural network. If the weight estimation model is a support vector machine,

The impression estimation model number ^ n is obtained, and the support vector machine is trained so that the impression estimation model number ^ n can be estimated by inputting the evaluator tendency q (i (j), k).

In step S132, the impression estimation model update unit 132 updates the impression estimation model n when the weight α (j, n) is given. The impression estimation model n is updated by the general method for each model, but the training weights of the samples x [j] and y [j] used for training each model are given by α (j, n) for learning. To do.

By alternately updating the weight estimation model update unit 131 and the impression estimation model update unit 132, the impression evaluator tendency group is automatically clustered into the impression estimation models 1 to N, and each evaluator's tendency is automatically clustered. The weight indicating which impression estimation model n is to be reflected as the estimation result is learned. At the time of estimation, an appropriate impression estimation unit 11-n is selected with a weight α according to the input evaluator tendency, and an integrated posterior probability for the input data is obtained.

In the present embodiment, the weight estimation model update unit 131 and the impression estimation model update unit 132 are updated alternately, but when both are modeled by the neural network, the weight estimation model update unit 131 and the weight estimation model are updated. Since the storage unit 16, the impression estimation model update unit 132, and the impression estimation model storage unit 111-n can be represented by one neural network, the error values between y (j) and y'(j) can be used at the same time. The weight estimation model and the impression estimation model can be updated.

<Impression estimation device>
As shown in FIG. 3, the impression estimation device 2 of the embodiment includes N (≧ 2) impression estimation units 11-n (n = 1, 2, ..., N), a weight estimation unit 12, and weight estimation. A model storage unit 16 and an estimated value integration unit 17 are provided. Each impression estimation unit 11-n includes an impression estimation model storage unit 111-n. The impression estimation device 2 realizes the impression estimation method of the embodiment by performing the processing of each step illustrated in FIG.

The impression estimation device 2 is configured by loading a special program into a known or dedicated computer having, for example, a central processing unit (CPU), a main storage device (RAM: Random Access Memory), or the like. Device. The impression estimation device 2 executes each process under the control of the central processing unit, for example. The data input to the impression estimation device 2 and the data obtained by each process are stored in, for example, the main storage device, and the data stored in the main storage device is read out to the central processing unit as needed. Used for other processing. At least a part of each processing unit of the impression estimation device 2 may be configured by hardware such as an integrated circuit. Each storage unit included in the impression estimation device 2 is, for example, a main storage device such as RAM (Random Access Memory), an auxiliary storage device composed of a hard disk, an optical disk, or a semiconductor memory element such as a flash memory, or an auxiliary storage device. It can be configured with middleware such as relational databases and key-value stores.

The impression estimation model storage unit 111-n stores the impression estimation model n learned by the impression estimation model learning device 1.

The weight estimation model storage unit 16 stores the weight estimation model learned by the impression estimation model learning device 1.

Hereinafter, the impression estimation method executed by the impression estimation device 2 of the embodiment will be described with reference to FIG.

In step S11, the impression estimation unit 11-n calculates and outputs the posterior probability p (n) of the impression value with respect to the input data x using the impression estimation model n stored in the impression estimation model storage unit 111-n. To do. The posterior probabilities p (n) of the impression values output by each impression estimation unit 11-n are input to the estimation value integration unit 17.

In step S12, the weight estimation unit 12 obtains the weight α (n) from the input evaluator tendency q (k) using the weight estimation model stored in the weight estimation model storage unit 16. The weight α (n) output by the weight estimation unit 12 is input to the estimation value integration unit 17.

The evaluator tendency q (k) input to the weight estimation unit 12 can give an arbitrary evaluator tendency. For example, if the evaluator tendency q (k) of a real evaluator is given, the impression value of the evaluator who has the same tendency as that evaluator can be estimated. Further, if the evaluator tendency q (k) of the virtual evaluator set manually is given, the impression value by the virtual evaluator can be simulated. Furthermore, it is also possible to obtain an average impression value by giving the average value of the evaluator tendency q (k) calculated from the entire evaluator.

In step S17, the estimation value integration unit 17 calculates the following equation using the posterior probabilities p (n) received from the impression estimation unit 11-n and the weight α (n) received from the weight estimation unit 12, and integrates them. Calculate the posterior probability y.

The estimated value integration unit 17 outputs the dimension of the highest posterior probability from the integrated posterior probability as an estimated impression value c. It is also possible to use the integrated posterior probability y itself as the impression value c.

In the present embodiment, the evaluator tendency q (k) of the evaluator is given as auxiliary information for impression estimation, but other information can also be given to the weight estimation unit 12 as auxiliary information. For example, predetermined environmental information may be given to the weight estimation unit 12 as auxiliary information in combination with the evaluator tendency q (k) of the evaluator. The environmental information is information about the environment that can affect the evaluation of the impression. For example, when the operator's impression (for example, favorability, brightness, etc.) is automatically estimated from the recorded voice of the call center, the call center is used. The evaluation of the impression may differ depending on the content of the reception (for example, failure reception, new reception, cancellation reception, customer consultation, etc.) and the skill of the call center staff (for example, primary reception staff, supervisor, etc.). A value that represents the reception content and skill attributes of is assumed. Further, in a scene where the impression of an image is evaluated, the brightness of the surroundings may affect the evaluation, so a value representing an attribute of the environment for evaluation such as indoor or outdoor is assumed. That is, by giving information on the environment that can affect the evaluation of the impression as auxiliary information, it is possible to estimate the impression that reflects the fluctuation of the evaluation due to the environment.

Although the embodiments of the present invention have been described above, the specific configuration is not limited to these embodiments, and even if the design is appropriately changed without departing from the spirit of the present invention, the specific configuration is not limited to these embodiments. Needless to say, it is included in the present invention. The various processes described in the embodiments are not only executed in chronological order according to the order described, but may also be executed in parallel or individually as required by the processing capacity of the device that executes the processes.

[Program, recording medium]
When various processing functions in each device described in the above embodiment are realized by a computer, the processing contents of the functions that each device should have are described by a program. Then, by executing this program on the computer, various processing functions in each of the above devices are realized on the computer.

The program describing the processing content can be recorded on a computer-readable recording medium. The computer-readable recording medium may be, for example, a magnetic recording device, an optical disk, a photomagnetic recording medium, a semiconductor memory, or the like.

In addition, the distribution of this program is carried out, for example, by selling, transferring, renting, or the like a portable recording medium such as a DVD or CD-ROM on which the program is recorded. Further, the program may be stored in the storage device of the server computer, and the program may be distributed by transferring the program from the server computer to another computer via a network.

A computer that executes such a program first, for example, first stores a program recorded on a portable recording medium or a program transferred from a server computer in its own storage device. Then, when the process is executed, the computer reads the program stored in its own storage device and executes the process according to the read program. Further, as another execution form of this program, a computer may read the program directly from a portable recording medium and execute processing according to the program, and further, the program is transferred from the server computer to this computer. Each time, the processing according to the received program may be executed sequentially. In addition, the above processing is executed by a so-called ASP (Application Service Provider) type service that realizes the processing function only by the execution instruction and result acquisition without transferring the program from the server computer to this computer. May be. The program in this embodiment includes information to be used for processing by a computer and equivalent to the program (data that is not a direct command to the computer but has a property of defining the processing of the computer, etc.).

Further, in this embodiment, the present device is configured by executing a predetermined program on the computer, but at least a part of these processing contents may be realized by hardware.

1 Impression estimation model learning device 2 Impression estimation device 11 Impression estimation unit 111 Impression estimation model storage unit 12 Weight estimation unit 13 Model learning unit 14 Learning data storage unit 15 Evaluator tendency data storage unit 16 Weight estimation model storage unit 17 Estimated value integration Department

Claims (8)

In order to obtain the impression value of the input data, a weight estimation model that estimates the weight from the evaluator tendency that represents the appearance probability of each impression value in the evaluator, and an impression estimation model that estimates the posterior probability of the impression value from the input data, It is an impression estimation model learning device that learns
The impression value of the input data is obtained based on the integrated posterior probability obtained by weighting the posterior probability of the impression value estimated by the impression estimation model with the weight estimated by the weight estimation model.
A learning data storage unit that stores a learning database that associates the learning data with the correct answer value of the impression value given to the learning data and the evaluator who gave the correct answer value of the impression value.
An evaluator tendency data storage unit that stores an evaluator tendency database that associates an evaluator with the above evaluator tendency,
And the impression estimation unit to estimate the posterior probability of the impression value from the learning data by using the impression estimation model in each group were classified on the Symbol evaluators trend,
A weight estimation unit that estimates the weight from the evaluator tendency associated with the learning data using the weight estimation model, and a weight estimation unit.
A weight estimation model update unit that updates the weight estimation model from the difference between the posterior probability of the impression value weighted by the weight and the correct answer value of the impression value given to the learning data.
An impression estimation model update unit that updates the impression estimation model using the weight as a learning weight from the posterior probability of the impression value,
Impression estimation model learning device including. The impression estimation model learning device according to claim 1.
The impression estimation unit determines the number of clusters based on the evaluator tendency database, and uses the impression estimation model for each group that classifies the evaluator tendency into the number of clusters. It estimates the probability,
Impression estimation model learning device. The impression estimation model learning device according to claim 1 or 2.
The evaluator tendency database associates the evaluator with the evaluator tendency and information on the environment in which the evaluation was performed.
The weight estimation model estimates the weights from at least information about the environment.
The weight estimation unit estimates the weight from at least the information about the environment associated with the learning data using the weight estimation model.
Impression estimation model learning device. An impression estimation model storage unit that stores an impression estimation model learned by the impression estimation model learning device according to any one of claims 1 to 3.
A weight estimation model storage unit that stores the weight estimation model learned by the impression estimation model learning device according to any one of claims 1 to 3.
An impression estimation unit that estimates the posterior probability of the impression value from the input data using the above impression estimation model for each group that classifies the evaluator tendency that represents the appearance probability of each impression value in the evaluator.
A weight estimation unit that estimates weights from the evaluator tendency input using the above weight estimation model,
An integrated value estimation unit that obtains the impression value of the input data based on the integrated posterior probability, which is the posterior probability of the impression value weighted by the weight.
Impression estimator including. The impression estimation device according to claim 4.
The integrated value estimation unit obtains the dimension having the highest value among the integrated posterior probabilities as an impression value of the input data.
Impression estimator. In order to obtain the impression value of the input data, a weight estimation model that estimates the weight from the evaluator tendency that represents the appearance probability of each impression value in the evaluator, and an impression estimation model that estimates the posterior probability of the impression value from the input data, It is an impression estimation model learning method to learn
The impression value of the input data is obtained based on the integrated posterior probability obtained by weighting the posterior probability of the impression value estimated by the impression estimation model with the weight estimated by the weight estimation model.
The learning data storage unit stores a learning database that associates the learning data with the correct answer value of the impression value given to the learning data and the evaluator who gave the correct answer value of the impression value.
The evaluator tendency data storage unit stores an evaluator tendency database that associates the evaluator with the above evaluator tendency.
Impression estimation section, using the impression estimation model to estimate the posterior probability of the impression value from the learning data for each group classified the above evaluation's trend,
The weight estimation unit estimates the weight from the evaluator tendency associated with the training data using the weight estimation model, and then estimates the weight.
The weight estimation model update unit updates the weight estimation model from the difference between the posterior probability of the impression value weighted by the weight and the correct answer value of the impression value given to the learning data.
The impression estimation model update unit updates the impression estimation model from the posterior probability of the impression value using the weight as a learning weight.
Impression estimation model learning method. The impression estimation model storage unit stores the impression estimation model learned by the impression estimation model learning method according to claim 6.
The weight estimation model stored in the weight estimation model storage unit stores the weight estimation model learned by the impression estimation model learning method according to claim 6.
The impression estimation unit estimates the posterior probability of the impression value from the input data using the above impression estimation model for each group that classifies the evaluator tendency representing the appearance probability of each impression value in the evaluator.
The weight estimation unit estimates the weight from the evaluator tendency input using the above weight estimation model.
The integrated value estimation unit obtains the impression value of the input data based on the integrated posterior probability, which is the posterior probability of the impression value weighted by the weight.
Impression estimation method. A program for operating a computer as the impression estimation model learning device according to any one of claims 1 to 3 or the impression estimation device according to claim 4 or 5.

Images